Anthropic Sat Out Huang's Open-Weights Letter

Jensen Huang's first post on X was not a product tease. It was a policy letter. On July 24 he published "Open Weights and American AI Leadership," signed at launch by Nvidia, Microsoft, Meta, IBM, and a stack of cloud, chip, and application companies. The ask was simple enough for a Capitol Hill one-pager: do not put premature restrictions on models whose parameters anyone can download, inspect, and run.

The roster then behaved like a live auction. Coverage noted OpenAI's absence. OpenAI signed. Coverage noted Google's absence. Google signed. By the next day Forbes and other outlets were counting 50 names, double the launch list, with AMD, Cisco, Cloudflare, GitHub, Block, and Ollama among the late arrivals. Distribution created the coalition. The signatures followed the attention.

4 min read

Anthropic did not move. Amazon did not move with it. Google, another Anthropic backer, signed anyway, which makes the Amazon pairing the more specific silence. Neither company offered a letter-specific explanation. Anthropic's public case against open weights has been consistent for years: once powerful parameters are released, you cannot readily revoke access, update every fork, or stop a copy that helps with cyberattacks or biological work.

That argument sits awkwardly beside Hugging Face's open-weight autopsy of the OpenAI agent intrusion. Defenders wanted a model they could host and bend. Safety labs want a model they can still kill. Both claims can be true about the same file. The policy fight is about which failure mode Washington fears more: a frontier weight that cannot be recalled, or a defender stuck waiting for a closed vendor's refusal stack to clear.

OpenAI's signature muddies the old closed-versus-open morality play. The company still warns about Chinese distillation and still sells proprietary access. Signing a letter that defends downloadable weights is a way to keep American open ecosystems alive while lobbying for targeted enforcement against alleged theft. It is also a way to stand next to Nvidia when the administration is eyeing Entity List tools aimed at Chinese labs.

The letter concedes the irreversibility problem and answers with an industry faith: enough outside researchers will find the bugs. That may be a good bet across years of collective scrutiny. It is a thin answer on the Tuesday a CIO has to prove which exact checkpoint signed a payment. Huang's coalition is arguing for openness in the layer where most signatories do not keep their deepest moat. CUDA remains closed. Azure still sells closed models. The weights are the layer they can afford to free.

Anthropic's empty slot is therefore the honest part of the week. A safety lab that believes released weights are ungovernable should refuse to applaud their proliferation, even when half of Silicon Valley is collecting signatures. The stampede can call that isolation. Isolation is sometimes the point.